swshon / dialectID_siam

Dialect identification using Siamese network

Dialect identification using Siamese network

Siamese network based dimensionality reduction for robust dialect identification (and as well as language recognition)

Requirment

• Python, tested on 2.7.6 (better with jupyter notebook)
• Tensorflow

Training Model

• train_ivector.ipynb : Training siamese neural network model for i-vector feature
• train_word.ipynb : Training siamese neural network model for word feature
• train_char.ipynb : Training siamese neural network model for character feature
• train_phone.ipynb : Training siamese neural network model for phoneme feature

Performance evaluation example of i-vector feature on MGB-3 Test dataset

• Confusion matrix

	EGY	GLF	LAB	MSA	NOR
EGY	225	13	37	11	24
GLF	11	178	31	16	12
LAB	43	45	212	14	31
MSA	5	5	10	208	10
NOR	18	9	44	13	267

• Precision

EGY	GLF	LAB	MSA	NOR
0.73	0.72	0.61	0.87	0.76

• Recall

EGY	GLF	LAB	MSA	NOR
0.75	0.71	0.63	0.79	0.78

Overall performance

Accurary : 0.731	Precision : 0.739	Recall : 0.732